* Merge from one-fftdataserver-per-fftmodel branch. This bit of reworking (which is not described very accurately by the title of the branch) turns the MatrixFile object into something that either reads or writes, but not both, and separates the FFT file cache reader and writer implementations separately. This allows the FFT data server to have a single thread owning writers and one reader per "customer" thread, and for all locking to be vastly simplified and concentrated in the data server alone (because none of the classes it makes use of is used in more than one thread at a time). The result is faster and more trustworthy code.